xAI finally shipped Grok 4.7 on Monday, September 21, 2026 — and "finally" is doing a lot of work in that sentence. Elon Musk had walked back the release timeline at least five times since late July: "four weeks out," then "a few weeks," then "3 to 4 weeks," then "10 days" on September 2, then "needs a few more days to cook" on September 11. When it actually landed, there was no staged rollout and no waitlist — it went live the same day in the Grok app, Cursor, Grok Build, and the xAI API.
The headline spec is size: Grok 4.7 runs on 2.1 trillion parameters, up roughly 40% from Grok 4.6's 1.5 trillion. But the more unusual detail is what xAI trained it on. Alongside the usual web-scale text, xAI folded in supplemental data pulled from SpaceX — Starlink satellite telemetry, manufacturing records, and engineering failure logs. The pitch is a model that reasons better about hardware and physical systems than one trained purely on internet text, which is a genuinely different data strategy than any of Grok's rivals are talking about.
Quick summary: Grok 4.7 launched September 21, 2026 with 2.1T parameters (up 40% from Grok 4.6), trained partly on SpaceX Starlink/manufacturing/engineering data. Pricing is unchanged at $2/$6 per million input/output tokens, and the context window stays at 500,000 tokens. It scores 46.3% on CursorBench 4.0 (vs. 40.4% for Grok 4.6) and 71.0% on DeepSWE v1.1 at high effort (vs. 65.2%). It arrives in an extremely crowded week: Claude Fable 5.1 (Sept 1), GPT-6 Astra (Sept 3-4), Claude Opus 5.5 and GPT-6 Sol/Luna (both Sept 22) all shipped within three weeks of it.
What actually changed
Three things moved with this release. First, scale: 2.1 trillion parameters versus Grok 4.6's 1.5 trillion. xAI hasn't published an architecture breakdown (active vs. total parameters for what's presumably a mixture-of-experts setup isn't public), so treat the raw parameter count as a headline number rather than a like-for-like comparison with dense models. Second, the SpaceX data injection — telemetry, manufacturing records, and failure logs that xAI says specifically improve reasoning about hardware and physical systems. Third, behavior: xAI says the model spends longer working through hard problems and double-checks its own answers more often than Grok 4.6 did, paired with what the company is calling its strongest safety guardrails yet.
On benchmarks, the gains show up most on longer-running work. CursorBench 4.0, which stresses extended coding sessions rather than single-shot snippets, moved from 40.4% (Grok 4.6) to 46.3% (Grok 4.7). On DeepSWE v1.1 at high effort, Grok 4.7 posts 71.0%, up from 65.2% for Grok 4.6. For context, GPT-6 Sol actually scores 68.8% on the same benchmark and Claude Fable 5.1 scores 67.4% — so Grok 4.7's 71.0% is ahead of both, though still behind GPT-6 Astra's reported 74.1%. Sources: Grok 4.7 Release: Same Price, Longer Horizons, Grok 4.7 vs Claude - Tech Insider.
What didn't change
Pricing held steady at $2 per million input tokens and $6 per million output tokens — the same rate as Grok 4.6, and still meaningfully cheaper than GPT-6 Astra's $10/$50 or Claude Fable 5.1's $10/$50. The context window also stayed flat at 500,000 tokens, which is less than half of GPT-6 Astra's 1.05 million or Fable 5.1's 1 million. xAI's framing is "a notable improvement over Grok 4.6 at the same price and speed" — an incremental-feeling release dressed up with a big parameter number and an interesting data story, rather than a leap on every axis.
The delay history, briefly
- Late July 2026: Musk says Grok 4.7 is "four weeks out."
- Early-to-mid August: timeline softens to "a few weeks," then "3 to 4 weeks."
- September 2, 2026: Musk says "10 days."
- September 11, 2026: Musk says the model "needs a few more days to cook."
- September 21, 2026: Grok 4.7 ships with no waitlist, live same-day across the Grok app, Cursor, Grok Build, and the API.
Landing in a very crowded three weeks
Grok 4.7's real problem isn't the model — it's the calendar. Anthropic shipped Claude Fable 5.1 (and the more restricted Claude Mythos 5.1) on September 1, 2026, with Fable 5.1 costing roughly 25% less than Fable 5 for typical workloads while improving on coding and long-running tasks. OpenAI shipped GPT-6 Astra to approved users on September 3, with general availability the next day, at $10/$50 per million tokens and a 1.05-million-token context window. Sources: Introducing Claude Fable 5.1 and Mythos 5.1 - Anthropic, GPT-6 Astra - OpenAI.
Then, in the 24 hours immediately after Grok 4.7 launched, both Anthropic and OpenAI moved again. Anthropic released Claude Opus 5.5 on September 22, 2026 — the first model in its new 5.5 family, priced at $4/$20 per million tokens (a 20% cut from Opus 5) with cache reads down 60% to $0.20 per million, and Anthropic says it performs near Fable 5.1 level on most work at 40% lower cost. The same day, OpenAI rolled out GPT-6 Sol and GPT-6 Luna: Sol targets complex coding work at $2/$10 per million tokens (down from GPT-5.6 Sol's $4/$20), while Luna handles high-volume clerical tasks like summarizing and extraction at $0.10/$0.50 per million tokens. Sources: Anthropic Releases Claude Opus 5.5 - NewsCord, OpenAI releases GPT-6 Sol and Luna - TechCrunch.
That's five significant model releases from three labs inside three weeks. Grok 4.7's SpaceX-flavored training data and its 2.1T parameter count are genuinely distinctive angles, but on raw benchmark and price comparisons it's landing in the middle of the pack rather than pulling ahead of it.
| Model | Release date | Parameters / context | Price (input/output per 1M tokens) | DeepSWE v1.1 (high effort) |
|---|---|---|---|---|
| Grok 4.7 | Sept 21, 2026 | 2.1T params, 500K context | $2 / $6 | 71.0% |
| Claude Fable 5.1 | Sept 1, 2026 | Undisclosed, 1M context | $10 / $50 | 67.4% |
| GPT-6 Astra | Sept 3-4, 2026 | Undisclosed, 1.05M context | $10 / $50 | 74.1% |
| GPT-6 Sol | Sept 22, 2026 | Undisclosed | $2 / $10 | 68.8% |
| Claude Opus 5.5 | Sept 22, 2026 | Undisclosed | $4 / $20 | Not directly comparable |
If you're picking a model for coding work today, don't anchor on parameter count alone. Grok 4.7's 2.1T is a headline figure with no disclosed active-parameter breakdown, and its DeepSWE score (71.0%) actually leads both GPT-6 Sol (68.8%) and Claude Fable 5.1 (67.4%) on that specific benchmark — though GPT-6 Astra (74.1%) tops all three. Run your own workload against two or three models before committing spend.
01How many parameters does Grok 4.7 have?
2.1 trillion, up about 40% from Grok 4.6's 1.5 trillion, according to xAI. The active-vs-total parameter split for its underlying architecture hasn't been disclosed.
02What's new about Grok 4.7's training data?
xAI added supplemental data from SpaceX, including Starlink satellite telemetry, manufacturing records, and engineering failure logs, aiming to improve reasoning about hardware and physical systems compared to models trained only on internet text.
03How much does Grok 4.7 cost?
$2 per million input tokens and $6 per million output tokens — unchanged from Grok 4.6, and cheaper than GPT-6 Astra or Claude Fable 5.1, both at $10/$50.
04Did the context window increase?
No. Grok 4.7 keeps the same 500,000-token context window as Grok 4.6, which is smaller than GPT-6 Astra's 1.05 million or Claude Fable 5.1's 1 million tokens.
05Why was Grok 4.7 delayed so many times?
xAI and Musk pushed the release back repeatedly since late July 2026 — from "four weeks out" to "a few weeks," "3 to 4 weeks," "10 days" (September 2), and "a few more days to cook" (September 11) — before it shipped without a waitlist on September 21.
06How does Grok 4.7 compare to GPT-6 Astra, Fable 5.1, Opus 5.5, and GPT-6 Sol/Luna?
It's cheaper than Astra, Fable 5.1, and Opus 5.5, and roughly matches GPT-6 Sol's price. On DeepSWE v1.1, it actually edges ahead of both — 71.0% versus Sol's 68.8% and Fable 5.1's 67.4% (GPT-6 Astra leads all three at 74.1%). Its edge is the SpaceX-derived training data for hardware and physical-systems reasoning, which none of the others claim.
With five frontier-ish models shipping inside three weeks, picking one by reading spec sheets alone is getting harder. LumiChats lets you compare and chat with Grok, GPT-6, Claude, and other leading models side by side in one place, so you can see how they actually handle your own prompts before deciding what to pay for.
